Sampling-based MOEAs

• Sampling-based MOEAs are evolutionary algorithms that integrate stochastic sampling in the environmental selection to allow sub-optimal solutions to persist.
• They enhance convergence on multi-objective benchmarks by preserving stepping-stone solutions, providing provable speedups over deterministic methods.
• Incorporating an external archive ensures retention of global Pareto-optimal solutions while enabling smaller populations for improved performance.

Sampling-based Multi-Objective Evolutionary Algorithms (SA_MOEA) employ stochastic mechanisms in the environmental selection phase, fundamentally altering how solution populations evolve across generations. Unlike the deterministic (greedy, elitist) selection systems that dominate classical MOEA frameworks, sampling-based approaches allow dominated or sub-optimal solutions to persist with nonzero probability. This stochasticity can substantially improve the ability to escape local optima and enables provable acceleration on certain benchmark problems, with both theoretical and empirical support.

1. Framework for Stochastic Population Update in MOEAs

The central mechanism of sampling-based MOEAs (hereafter SA_MOEA, Editor's term) is the stochastic modification of population update. In each generation, instead of deterministically selecting the “worst” solution for removal from the entire union of parents and offspring, SA_MOEA stochastically samples a random subset from this pool and applies the standard selection operator (hypervolume-loss or crowding distance) only within this subset.

Generic SA_MOEA Update Pattern

For a pool $Q$ of candidate solutions (parent population $P$ and offspring set $O$):

1. Draw a random subset $Q'$ of predetermined size from $Q$, uniformly at random without replacement.
2. Apply the environmental selection operator within $Q'$ only (hypervolume loss for SMS-EMOA, non-dominated sorting and crowding distance for NSGA-II).
3. Delete the chosen individual(s) from $Q$ based solely on competition within $Q'$, retaining all others.

The only departure from traditional update is the random sampling step. In typical instantiations:

• $|Q'| = \lfloor |Q|/2 \rfloor$ (SMS-EMOA)
• $|Q'| = \lfloor 3|Q|/4 \rfloor$ (NSGA-II)

This mechanism guarantees every solution in $P$0 a probability $P$1 of being “in competition” during the selection step (Bian et al., 2023).

2. Technical Instantiations: SMS-EMOA and NSGA-II

The stochastic population update framework was operationalized in both SMS-EMOA and NSGA-II in (Bian et al., 2023).

SMS-EMOA (Stochastic Update)

• Population $P$2 (size $P$3), offspring $P$4 $P$5
• Randomly sample $P$6 ($P$7)
• Partition $P$8 into non-dominated fronts $P$9
• Remove $O$0 where $O$1 is hypervolume loss
• Next population: $O$2

NSGA-II (Stochastic Update)

• Population $O$3 (size $O$4), offspring $O$5 (size $O$6), $O$7
• Randomly sample $O$8 ($O$9)
• Partition $Q'$0 by non-dominated fronts $Q'$1; greedily fill next population by fronts; use crowding distance to break ties in critical front
• Survivors: $Q'$2, where $Q'$3 are the selected solutions from $Q'$4

Selection and mutation probabilities are precisely quantified. For any $Q'$5, the survival (competition) probability is $Q'$6. Mutation (bit-flip) events follow standard binomial event probabilities (Bian et al., 2023).

3. Theoretical Analysis and Performance Guarantees

Analytic and proof-based assessment of SA_MOEA has focused on specifically-constructed multi-objective benchmarks, including OneJumpZeroJump (OJZJ) and the bi-objective RealRoyalRoad (RRMO):

OneJumpZeroJump (OJZJ):

• Binary string $Q'$7 of length $Q'$8, parameter $Q'$9
• Objectives:
  • $Q$0 if $Q$1 or $Q$2, else $Q$3
  • $Q$4 if $Q$5 or $Q$6, else $Q$7
• Pareto front size = $Q$8

RealRoyalRoad (RRMO):

• Binary string $Q$9
• Partitioned regions and objectives based on counts of 1s and specialized substring functions LZ, TZ

Main Theoretical Results

The core finding: stochastic population update provably accelerates Pareto front discovery relative to deterministic update, sometimes exponentially.

• For SMS-EMOA on OJZJ ($Q'$0):
  • Deterministic: $Q'$1, lower bound $Q'$2
  • Stochastic: $Q'$3
  • With $Q'$4, this yields exponential improvement by factor $Q'$5
• For NSGA-II:
  • OJZJ: deterministic $Q'$6; stochastic $Q'$7
  • RRMO: deterministic $Q'$8; stochastic $Q'$9

Mechanism: In deterministic MOEA, the search must traverse large valleys of dominated solutions, requiring rare high-order improvements. Stochastic update ensures that dominated but “helpful” stepping-stone solutions survive with constant probability, facilitating incremental progress by single-bit (or other local) refinements (Bian et al., 2023).

4. Critique and Enhancement: Need for an External Archive

While the SA_MOEA mechanism enables beneficial non-elitism, one drawback is the potential loss of globally Pareto-optimal solutions from the evolving population. This limitation imposes a requirement for larger populations to ensure elite solutions are retained, which in turn diminishes the effect of sampling-based selection.

The introduction of an unbounded external archive resolves this problem (Ren et al., 28 Jan 2025). The archive stores every non-dominated solution ever found. After each offspring is generated, if it is not strictly dominated by any archive member, it is inserted, and any members weakly dominated by it are removed. This mechanism:

• Guarantees global Pareto-optimal points are never lost, regardless of stochastic population effects
• Allows for greatly reduced in-population size ($Q$0) without risking loss of elite solutions
• Amplifies the effective survival rate of dominated but promising solutions, increasing the speed-up effect

Empirically and theoretically, coupling SA_MOEA with an archive yields best-in-class performance, particularly on challenging benchmarks and practical problems such as the multi-objective traveling salesperson problem (Ren et al., 28 Jan 2025).

5. Empirical Results and Benchmark Performance

The SA_MOEA framework, both with and without archives, has been empirically validated on canonical benchmarks:

OneJumpZeroJump ($Q$1):

• SMS-EMOA with stochastic update: $Q$2–$Q$3 reduction in generations to full Pareto cover versus deterministic
• NSGA-II: $Q$4–$Q$5 speed-up observed

RealRoyalRoad:

• Speedup ratios (deterministic/stochastic) increase nearly exponentially with $Q$6
  • For $Q$7: SMS-EMOA, $Q$8 fewer generations; NSGA-II, $Q$9 fewer generations
  • SMS-EMOA at $Q'$0: $Q'$1 (deterministic) vs $Q'$2 (stochastic)
  • NSGA-II at $Q'$3: $Q'$4 vs $Q'$5

External-archive variants in subsequent work consistently outperform non-archive versions, particularly in maintaining optimal coverage with much smaller $Q'$6, further reinforcing provable speedups (Bian et al., 2023, Ren et al., 28 Jan 2025).

6. Relation to Broader Sampling-based Strategies in EAs

SA_MOEA’s core principle aligns with broader trends in evolutionary optimization leveraging uncertainty and probability to escape local optima—such as in surrogate-assisted and probabilistic selection mechanisms for constrained or expensive indicator problems. For instance, probabilistic selection-based EAs like PSCMOEA use surrogate uncertainty to probabilistically rank and select candidates under constraints, blending model-driven sampling with adaptive exploration (Rahi, 2024). However, in the case of SA_MOEA, the randomness is directly embedded into the population update, rather than via model-based sampling.

7. Algorithmic Considerations and Practical Recommendations

Stochastic update is best combined with the $Q'$7-type update of NSGA-II, rather than the $Q'$8-mode of SMS-EMOA, as proven in (Ren et al., 28 Jan 2025). For SMS-EMOA, the effectiveness of the technique is highly sensitive to population size; archival support is necessary to keep $Q'$9 small and sampling rates high. Empirical results on both synthetic and real-world problems demonstrate that:

• Archive plus stochastic update (“SPU+archive”) consistently outperforms deterministic, stochastic-only, and archive-only baselines
• The improvement is most pronounced where combinatorial paths to the Pareto front are long and “stepping stone” survival is critical

In conclusion, sampling-based MOEAs introduce a minimal but fundamentally advantageous change to environmental selection in evolutionary optimization. By harnessing stochasticity to allow temporary persistence of inferior solutions, they enable both theoretically provable and empirically validated exponential speedups in challenging multi-objective landscapes, especially when external archiving is employed to capture and persist global optima (Bian et al., 2023, Ren et al., 28 Jan 2025).